services/keep-web/s3.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package keepweb
   6
   7 import (
   8         "crypto/hmac"
   9         "crypto/sha256"
  10         "encoding/base64"
  11         "encoding/json"
  12         "encoding/xml"
  13         "errors"
  14         "fmt"
  15         "hash"
  16         "io"
  17         "mime"
  18         "net/http"
  19         "net/textproto"
  20         "net/url"
  21         "os"
  22         "path/filepath"
  23         "regexp"
  24         "sort"
  25         "strconv"
  26         "strings"
  27         "time"
  28
  29         "git.arvados.org/arvados.git/sdk/go/arvados"
  30         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  31         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  32         "git.arvados.org/arvados.git/sdk/go/keepclient"
  33         "github.com/AdRoll/goamz/s3"
  34 )
  35
  36 const (
  37         s3MaxKeys       = 1000
  38         s3SignAlgorithm = "AWS4-HMAC-SHA256"
  39         s3MaxClockSkew  = 5 * time.Minute
  40 )
  41
  42 type commonPrefix struct {
  43         Prefix string
  44 }
  45
  46 type listV1Resp struct {
  47         XMLName string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  48         s3.ListResp
  49         // s3.ListResp marshals an empty tag when
  50         // CommonPrefixes is nil, which confuses some clients.
  51         // Fix by using this nested struct instead.
  52         CommonPrefixes []commonPrefix
  53         // Similarly, we need omitempty here, because an empty
  54         // tag confuses some clients (e.g.,
  55         // github.com/aws/aws-sdk-net never terminates its
  56         // paging loop).
  57         NextMarker string `xml:"NextMarker,omitempty"`
  58         // ListObjectsV2 has a KeyCount response field.
  59         KeyCount int
  60 }
  61
  62 type listV2Resp struct {
  63         XMLName               string `xml:"http://s3.amazonaws.com/doc/2006-03-01/ ListBucketResult"`
  64         IsTruncated           bool
  65         Contents              []s3.Key
  66         Name                  string
  67         Prefix                string
  68         Delimiter             string
  69         MaxKeys               int
  70         CommonPrefixes        []commonPrefix
  71         EncodingType          string `xml:",omitempty"`
  72         KeyCount              int
  73         ContinuationToken     string `xml:",omitempty"`
  74         NextContinuationToken string `xml:",omitempty"`
  75         StartAfter            string `xml:",omitempty"`
  76 }
  77
  78 func hmacstring(msg string, key []byte) []byte {
  79         h := hmac.New(sha256.New, key)
  80         io.WriteString(h, msg)
  81         return h.Sum(nil)
  82 }
  83
  84 func hashdigest(h hash.Hash, payload string) string {
  85         io.WriteString(h, payload)
  86         return fmt.Sprintf("%x", h.Sum(nil))
  87 }
  88
  89 // Signing key for given secret key and request attrs.
  90 func s3signatureKey(key, datestamp, regionName, serviceName string) []byte {
  91         return hmacstring("aws4_request",
  92                 hmacstring(serviceName,
  93                         hmacstring(regionName,
  94                                 hmacstring(datestamp, []byte("AWS4"+key)))))
  95 }
  96
  97 // Canonical query string for S3 V4 signature: sorted keys, spaces
  98 // escaped as %20 instead of +, keyvalues joined with &.
  99 func s3querystring(u *url.URL) string {
 100         keys := make([]string, 0, len(u.Query()))
 101         values := make(map[string]string, len(u.Query()))
 102         for k, vs := range u.Query() {
 103                 k = strings.Replace(url.QueryEscape(k), "+", "%20", -1)
 104                 keys = append(keys, k)
 105                 for _, v := range vs {
 106                         v = strings.Replace(url.QueryEscape(v), "+", "%20", -1)
 107                         if values[k] != "" {
 108                                 values[k] += "&"
 109                         }
 110                         values[k] += k + "=" + v
 111                 }
 112         }
 113         sort.Strings(keys)
 114         for i, k := range keys {
 115                 keys[i] = values[k]
 116         }
 117         return strings.Join(keys, "&")
 118 }
 119
 120 var reMultipleSlashChars = regexp.MustCompile(`//+`)
 121
 122 func s3stringToSign(alg, scope, signedHeaders string, r *http.Request) (string, error) {
 123         timefmt, timestr := "20060102T150405Z", r.Header.Get("X-Amz-Date")
 124         if timestr == "" {
 125                 timefmt, timestr = time.RFC1123, r.Header.Get("Date")
 126         }
 127         t, err := time.Parse(timefmt, timestr)
 128         if err != nil {
 129                 return "", fmt.Errorf("invalid timestamp %q: %s", timestr, err)
 130         }
 131         if skew := time.Now().Sub(t); skew < -s3MaxClockSkew || skew > s3MaxClockSkew {
 132                 return "", errors.New("exceeded max clock skew")
 133         }
 134
 135         var canonicalHeaders string
 136         for _, h := range strings.Split(signedHeaders, ";") {
 137                 if h == "host" {
 138                         canonicalHeaders += h + ":" + r.Host + "\n"
 139                 } else {
 140                         canonicalHeaders += h + ":" + r.Header.Get(h) + "\n"
 141                 }
 142         }
 143
 144         normalizedPath := normalizePath(r.URL.Path)
 145         ctxlog.FromContext(r.Context()).Debugf("normalizedPath %q", normalizedPath)
 146         canonicalRequest := fmt.Sprintf("%s\n%s\n%s\n%s\n%s\n%s", r.Method, normalizedPath, s3querystring(r.URL), canonicalHeaders, signedHeaders, r.Header.Get("X-Amz-Content-Sha256"))
 147         ctxlog.FromContext(r.Context()).Debugf("s3stringToSign: canonicalRequest %s", canonicalRequest)
 148         return fmt.Sprintf("%s\n%s\n%s\n%s", alg, r.Header.Get("X-Amz-Date"), scope, hashdigest(sha256.New(), canonicalRequest)), nil
 149 }
 150
 151 func normalizePath(s string) string {
 152         // (url.URL).EscapedPath() would be incorrect here. AWS
 153         // documentation specifies the URL path should be normalized
 154         // according to RFC 3986, i.e., unescaping ALPHA / DIGIT / "-"
 155         // / "." / "_" / "~". The implication is that everything other
 156         // than those chars (and "/") _must_ be percent-encoded --
 157         // even chars like ";" and "," that are not normally
 158         // percent-encoded in paths.
 159         out := ""
 160         for _, c := range []byte(reMultipleSlashChars.ReplaceAllString(s, "/")) {
 161                 if (c >= 'a' && c <= 'z') ||
 162                         (c >= 'A' && c <= 'Z') ||
 163                         (c >= '0' && c <= '9') ||
 164                         c == '-' ||
 165                         c == '.' ||
 166                         c == '_' ||
 167                         c == '~' ||
 168                         c == '/' {
 169                         out += string(c)
 170                 } else {
 171                         out += fmt.Sprintf("%%%02X", c)
 172                 }
 173         }
 174         return out
 175 }
 176
 177 func s3signature(secretKey, scope, signedHeaders, stringToSign string) (string, error) {
 178         // scope is {datestamp}/{region}/{service}/aws4_request
 179         drs := strings.Split(scope, "/")
 180         if len(drs) != 4 {
 181                 return "", fmt.Errorf("invalid scope %q", scope)
 182         }
 183         key := s3signatureKey(secretKey, drs[0], drs[1], drs[2])
 184         return hashdigest(hmac.New(sha256.New, key), stringToSign), nil
 185 }
 186
 187 var v2tokenUnderscore = regexp.MustCompile(`^v2_[a-z0-9]{5}-gj3su-[a-z0-9]{15}_`)
 188
 189 func unescapeKey(key string) string {
 190         if v2tokenUnderscore.MatchString(key) {
 191                 // Entire Arvados token, with "/" replaced by "_" to
 192                 // avoid colliding with the Authorization header
 193                 // format.
 194                 return strings.Replace(key, "_", "/", -1)
 195         } else if s, err := url.PathUnescape(key); err == nil {
 196                 return s
 197         } else {
 198                 return key
 199         }
 200 }
 201
 202 // checks3signature verifies the given S3 V4 signature and returns the
 203 // Arvados token that corresponds to the given accessKey. An error is
 204 // returned if accessKey is not a valid token UUID or the signature
 205 // does not match.
 206 func (h *handler) checks3signature(r *http.Request) (string, error) {
 207         var key, scope, signedHeaders, signature string
 208         authstring := strings.TrimPrefix(r.Header.Get("Authorization"), s3SignAlgorithm+" ")
 209         for _, cmpt := range strings.Split(authstring, ",") {
 210                 cmpt = strings.TrimSpace(cmpt)
 211                 split := strings.SplitN(cmpt, "=", 2)
 212                 switch {
 213                 case len(split) != 2:
 214                         // (?) ignore
 215                 case split[0] == "Credential":
 216                         keyandscope := strings.SplitN(split[1], "/", 2)
 217                         if len(keyandscope) == 2 {
 218                                 key, scope = keyandscope[0], keyandscope[1]
 219                         }
 220                 case split[0] == "SignedHeaders":
 221                         signedHeaders = split[1]
 222                 case split[0] == "Signature":
 223                         signature = split[1]
 224                 }
 225         }
 226
 227         client := (&arvados.Client{
 228                 APIHost:  h.Cluster.Services.Controller.ExternalURL.Host,
 229                 Insecure: h.Cluster.TLS.Insecure,
 230         }).WithRequestID(r.Header.Get("X-Request-Id"))
 231         var aca arvados.APIClientAuthorization
 232         var secret string
 233         var err error
 234         if len(key) == 27 && key[5:12] == "-gj3su-" {
 235                 // Access key is the UUID of an Arvados token, secret
 236                 // key is the secret part.
 237                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+h.Cluster.SystemRootToken)
 238                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/"+key, nil, nil)
 239                 secret = aca.APIToken
 240         } else {
 241                 // Access key and secret key are both an entire
 242                 // Arvados token or OIDC access token.
 243                 ctx := arvados.ContextWithAuthorization(r.Context(), "Bearer "+unescapeKey(key))
 244                 err = client.RequestAndDecodeContext(ctx, &aca, "GET", "arvados/v1/api_client_authorizations/current", nil, nil)
 245                 secret = key
 246         }
 247         if err != nil {
 248                 ctxlog.FromContext(r.Context()).WithError(err).WithField("UUID", key).Info("token lookup failed")
 249                 return "", errors.New("invalid access key")
 250         }
 251         stringToSign, err := s3stringToSign(s3SignAlgorithm, scope, signedHeaders, r)
 252         if err != nil {
 253                 return "", err
 254         }
 255         expect, err := s3signature(secret, scope, signedHeaders, stringToSign)
 256         if err != nil {
 257                 return "", err
 258         } else if expect != signature {
 259                 return "", fmt.Errorf("signature does not match (scope %q signedHeaders %q stringToSign %q)", scope, signedHeaders, stringToSign)
 260         }
 261         return aca.TokenV2(), nil
 262 }
 263
 264 func s3ErrorResponse(w http.ResponseWriter, s3code string, message string, resource string, code int) {
 265         w.Header().Set("Content-Type", "application/xml")
 266         w.Header().Set("X-Content-Type-Options", "nosniff")
 267         w.WriteHeader(code)
 268         var errstruct struct {
 269                 Code      string
 270                 Message   string
 271                 Resource  string
 272                 RequestId string
 273         }
 274         errstruct.Code = s3code
 275         errstruct.Message = message
 276         errstruct.Resource = resource
 277         errstruct.RequestId = ""
 278         enc := xml.NewEncoder(w)
 279         fmt.Fprint(w, xml.Header)
 280         enc.EncodeElement(errstruct, xml.StartElement{Name: xml.Name{Local: "Error"}})
 281 }
 282
 283 var NoSuchKey = "NoSuchKey"
 284 var NoSuchBucket = "NoSuchBucket"
 285 var InvalidArgument = "InvalidArgument"
 286 var InternalError = "InternalError"
 287 var UnauthorizedAccess = "UnauthorizedAccess"
 288 var InvalidRequest = "InvalidRequest"
 289 var SignatureDoesNotMatch = "SignatureDoesNotMatch"
 290
 291 var reRawQueryIndicatesAPI = regexp.MustCompile(`^[a-z]+(&|$)`)
 292
 293 // serveS3 handles r and returns true if r is a request from an S3
 294 // client, otherwise it returns false.
 295 func (h *handler) serveS3(w http.ResponseWriter, r *http.Request) bool {
 296         var token string
 297         if auth := r.Header.Get("Authorization"); strings.HasPrefix(auth, "AWS ") {
 298                 split := strings.SplitN(auth[4:], ":", 2)
 299                 if len(split) < 2 {
 300                         s3ErrorResponse(w, InvalidRequest, "malformed Authorization header", r.URL.Path, http.StatusUnauthorized)
 301                         return true
 302                 }
 303                 token = unescapeKey(split[0])
 304         } else if strings.HasPrefix(auth, s3SignAlgorithm+" ") {
 305                 t, err := h.checks3signature(r)
 306                 if err != nil {
 307                         s3ErrorResponse(w, SignatureDoesNotMatch, "signature verification failed: "+err.Error(), r.URL.Path, http.StatusForbidden)
 308                         return true
 309                 }
 310                 token = t
 311         } else {
 312                 return false
 313         }
 314
 315         var err error
 316         var fs arvados.CustomFileSystem
 317         var arvclient *arvadosclient.ArvadosClient
 318         if r.Method == http.MethodGet || r.Method == http.MethodHead {
 319                 // Use a single session (cached FileSystem) across
 320                 // multiple read requests.
 321                 var sess *cachedSession
 322                 fs, sess, err = h.Cache.GetSession(token)
 323                 if err != nil {
 324                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 325                         return true
 326                 }
 327                 arvclient = sess.arvadosclient
 328         } else {
 329                 // Create a FileSystem for this request, to avoid
 330                 // exposing incomplete write operations to concurrent
 331                 // requests.
 332                 var kc *keepclient.KeepClient
 333                 var release func()
 334                 var client *arvados.Client
 335                 arvclient, kc, client, release, err = h.getClients(r.Header.Get("X-Request-Id"), token)
 336                 if err != nil {
 337                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 338                         return true
 339                 }
 340                 defer release()
 341                 fs = client.SiteFileSystem(kc)
 342                 fs.ForwardSlashNameSubstitution(h.Cluster.Collections.ForwardSlashNameSubstitution)
 343         }
 344
 345         var objectNameGiven bool
 346         var bucketName string
 347         fspath := "/by_id"
 348         if id := arvados.CollectionIDFromDNSName(r.Host); id != "" {
 349                 fspath += "/" + id
 350                 bucketName = id
 351                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 0
 352         } else {
 353                 bucketName = strings.SplitN(strings.TrimPrefix(r.URL.Path, "/"), "/", 2)[0]
 354                 objectNameGiven = strings.Count(strings.TrimSuffix(r.URL.Path, "/"), "/") > 1
 355         }
 356         fspath += reMultipleSlashChars.ReplaceAllString(r.URL.Path, "/")
 357
 358         switch {
 359         case r.Method == http.MethodGet && !objectNameGiven:
 360                 // Path is "/{uuid}" or "/{uuid}/", has no object name
 361                 if _, ok := r.URL.Query()["versioning"]; ok {
 362                         // GetBucketVersioning
 363                         w.Header().Set("Content-Type", "application/xml")
 364                         io.WriteString(w, xml.Header)
 365                         fmt.Fprintln(w, `<VersioningConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/"/>`)
 366                 } else if _, ok = r.URL.Query()["location"]; ok {
 367                         // GetBucketLocation
 368                         w.Header().Set("Content-Type", "application/xml")
 369                         io.WriteString(w, xml.Header)
 370                         fmt.Fprintln(w, `<LocationConstraint><LocationConstraint xmlns="http://s3.amazonaws.com/doc/2006-03-01/">`+
 371                                 h.Cluster.ClusterID+
 372                                 `</LocationConstraint></LocationConstraint>`)
 373                 } else if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 374                         // GetBucketWebsite ("GET /bucketid/?website"), GetBucketTagging, etc.
 375                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 376                 } else {
 377                         // ListObjects
 378                         h.s3list(bucketName, w, r, fs)
 379                 }
 380                 return true
 381         case r.Method == http.MethodGet || r.Method == http.MethodHead:
 382                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 383                         // GetObjectRetention ("GET /bucketid/objectid?retention&versionID=..."), etc.
 384                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 385                         return true
 386                 }
 387                 fi, err := fs.Stat(fspath)
 388                 if r.Method == "HEAD" && !objectNameGiven {
 389                         // HeadBucket
 390                         if err == nil && fi.IsDir() {
 391                                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 392                                 if err != nil {
 393                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 394                                         return true
 395                                 }
 396                                 w.WriteHeader(http.StatusOK)
 397                         } else if os.IsNotExist(err) {
 398                                 s3ErrorResponse(w, NoSuchBucket, "The specified bucket does not exist.", r.URL.Path, http.StatusNotFound)
 399                         } else {
 400                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 401                         }
 402                         return true
 403                 }
 404                 if err == nil && fi.IsDir() && objectNameGiven && strings.HasSuffix(fspath, "/") && h.Cluster.Collections.S3FolderObjects {
 405                         err = setFileInfoHeaders(w.Header(), fs, fspath)
 406                         if err != nil {
 407                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 408                                 return true
 409                         }
 410                         w.Header().Set("Content-Type", "application/x-directory")
 411                         w.WriteHeader(http.StatusOK)
 412                         return true
 413                 }
 414                 if os.IsNotExist(err) ||
 415                         (err != nil && err.Error() == "not a directory") ||
 416                         (fi != nil && fi.IsDir()) {
 417                         s3ErrorResponse(w, NoSuchKey, "The specified key does not exist.", r.URL.Path, http.StatusNotFound)
 418                         return true
 419                 }
 420
 421                 tokenUser, err := h.Cache.GetTokenUser(token)
 422                 if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 423                         http.Error(w, "Not permitted", http.StatusForbidden)
 424                         return true
 425                 }
 426                 h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 427
 428                 // shallow copy r, and change URL path
 429                 r := *r
 430                 r.URL.Path = fspath
 431                 err = setFileInfoHeaders(w.Header(), fs, fspath)
 432                 if err != nil {
 433                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 434                         return true
 435                 }
 436                 http.FileServer(fs).ServeHTTP(w, &r)
 437                 return true
 438         case r.Method == http.MethodPut:
 439                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 440                         // PutObjectAcl ("PUT /bucketid/objectid?acl&versionID=..."), etc.
 441                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 442                         return true
 443                 }
 444                 if !objectNameGiven {
 445                         s3ErrorResponse(w, InvalidArgument, "Missing object name in PUT request.", r.URL.Path, http.StatusBadRequest)
 446                         return true
 447                 }
 448                 var objectIsDir bool
 449                 if strings.HasSuffix(fspath, "/") {
 450                         if !h.Cluster.Collections.S3FolderObjects {
 451                                 s3ErrorResponse(w, InvalidArgument, "invalid object name: trailing slash", r.URL.Path, http.StatusBadRequest)
 452                                 return true
 453                         }
 454                         n, err := r.Body.Read(make([]byte, 1))
 455                         if err != nil && err != io.EOF {
 456                                 s3ErrorResponse(w, InternalError, fmt.Sprintf("error reading request body: %s", err), r.URL.Path, http.StatusInternalServerError)
 457                                 return true
 458                         } else if n > 0 {
 459                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless content is empty", r.URL.Path, http.StatusBadRequest)
 460                                 return true
 461                         } else if strings.SplitN(r.Header.Get("Content-Type"), ";", 2)[0] != "application/x-directory" {
 462                                 s3ErrorResponse(w, InvalidArgument, "cannot create object with trailing '/' char unless Content-Type is 'application/x-directory'", r.URL.Path, http.StatusBadRequest)
 463                                 return true
 464                         }
 465                         // Given PUT "foo/bar/", we'll use "foo/bar/."
 466                         // in the "ensure parents exist" block below,
 467                         // and then we'll be done.
 468                         fspath += "."
 469                         objectIsDir = true
 470                 }
 471                 fi, err := fs.Stat(fspath)
 472                 if err != nil && err.Error() == "not a directory" {
 473                         // requested foo/bar, but foo is a file
 474                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 475                         return true
 476                 }
 477                 if strings.HasSuffix(r.URL.Path, "/") && err == nil && !fi.IsDir() {
 478                         // requested foo/bar/, but foo/bar is a file
 479                         s3ErrorResponse(w, InvalidArgument, "object name conflicts with existing object", r.URL.Path, http.StatusBadRequest)
 480                         return true
 481                 }
 482                 // create missing parent/intermediate directories, if any
 483                 for i, c := range fspath {
 484                         if i > 0 && c == '/' {
 485                                 dir := fspath[:i]
 486                                 if strings.HasSuffix(dir, "/") {
 487                                         err = errors.New("invalid object name (consecutive '/' chars)")
 488                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 489                                         return true
 490                                 }
 491                                 err = fs.Mkdir(dir, 0755)
 492                                 if errors.Is(err, arvados.ErrInvalidArgument) || errors.Is(err, arvados.ErrInvalidOperation) {
 493                                         // Cannot create a directory
 494                                         // here.
 495                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 496                                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 497                                         return true
 498                                 } else if err != nil && !os.IsExist(err) {
 499                                         err = fmt.Errorf("mkdir %q failed: %w", dir, err)
 500                                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 501                                         return true
 502                                 }
 503                         }
 504                 }
 505                 if !objectIsDir {
 506                         f, err := fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 507                         if os.IsNotExist(err) {
 508                                 f, err = fs.OpenFile(fspath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
 509                         }
 510                         if err != nil {
 511                                 err = fmt.Errorf("open %q failed: %w", r.URL.Path, err)
 512                                 s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 513                                 return true
 514                         }
 515                         defer f.Close()
 516
 517                         tokenUser, err := h.Cache.GetTokenUser(token)
 518                         if !h.userPermittedToUploadOrDownload(r.Method, tokenUser) {
 519                                 http.Error(w, "Not permitted", http.StatusForbidden)
 520                                 return true
 521                         }
 522                         h.logUploadOrDownload(r, arvclient, fs, fspath, nil, tokenUser)
 523
 524                         _, err = io.Copy(f, r.Body)
 525                         if err != nil {
 526                                 err = fmt.Errorf("write to %q failed: %w", r.URL.Path, err)
 527                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 528                                 return true
 529                         }
 530                         err = f.Close()
 531                         if err != nil {
 532                                 err = fmt.Errorf("write to %q failed: close: %w", r.URL.Path, err)
 533                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusBadGateway)
 534                                 return true
 535                         }
 536                 }
 537                 err = fs.Sync()
 538                 if err != nil {
 539                         err = fmt.Errorf("sync failed: %w", err)
 540                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 541                         return true
 542                 }
 543                 // Ensure a subsequent read operation will see the changes.
 544                 h.Cache.ResetSession(token)
 545                 w.WriteHeader(http.StatusOK)
 546                 return true
 547         case r.Method == http.MethodDelete:
 548                 if reRawQueryIndicatesAPI.MatchString(r.URL.RawQuery) {
 549                         // DeleteObjectTagging ("DELETE /bucketid/objectid?tagging&versionID=..."), etc.
 550                         s3ErrorResponse(w, InvalidRequest, "API not supported", r.URL.Path+"?"+r.URL.RawQuery, http.StatusBadRequest)
 551                         return true
 552                 }
 553                 if !objectNameGiven || r.URL.Path == "/" {
 554                         s3ErrorResponse(w, InvalidArgument, "missing object name in DELETE request", r.URL.Path, http.StatusBadRequest)
 555                         return true
 556                 }
 557                 if strings.HasSuffix(fspath, "/") {
 558                         fspath = strings.TrimSuffix(fspath, "/")
 559                         fi, err := fs.Stat(fspath)
 560                         if os.IsNotExist(err) {
 561                                 w.WriteHeader(http.StatusNoContent)
 562                                 return true
 563                         } else if err != nil {
 564                                 s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 565                                 return true
 566                         } else if !fi.IsDir() {
 567                                 // if "foo" exists and is a file, then
 568                                 // "foo/" doesn't exist, so we say
 569                                 // delete was successful.
 570                                 w.WriteHeader(http.StatusNoContent)
 571                                 return true
 572                         }
 573                 } else if fi, err := fs.Stat(fspath); err == nil && fi.IsDir() {
 574                         // if "foo" is a dir, it is visible via S3
 575                         // only as "foo/", not "foo" -- so we leave
 576                         // the dir alone and return 204 to indicate
 577                         // that "foo" does not exist.
 578                         w.WriteHeader(http.StatusNoContent)
 579                         return true
 580                 }
 581                 err = fs.Remove(fspath)
 582                 if os.IsNotExist(err) {
 583                         w.WriteHeader(http.StatusNoContent)
 584                         return true
 585                 }
 586                 if err != nil {
 587                         err = fmt.Errorf("rm failed: %w", err)
 588                         s3ErrorResponse(w, InvalidArgument, err.Error(), r.URL.Path, http.StatusBadRequest)
 589                         return true
 590                 }
 591                 err = fs.Sync()
 592                 if err != nil {
 593                         err = fmt.Errorf("sync failed: %w", err)
 594                         s3ErrorResponse(w, InternalError, err.Error(), r.URL.Path, http.StatusInternalServerError)
 595                         return true
 596                 }
 597                 // Ensure a subsequent read operation will see the changes.
 598                 h.Cache.ResetSession(token)
 599                 w.WriteHeader(http.StatusNoContent)
 600                 return true
 601         default:
 602                 s3ErrorResponse(w, InvalidRequest, "method not allowed", r.URL.Path, http.StatusMethodNotAllowed)
 603                 return true
 604         }
 605 }
 606
 607 func setFileInfoHeaders(header http.Header, fs arvados.CustomFileSystem, path string) error {
 608         maybeEncode := func(s string) string {
 609                 for _, c := range s {
 610                         if c > '\u007f' {
 611                                 return mime.BEncoding.Encode("UTF-8", s)
 612                         }
 613                 }
 614                 return s
 615         }
 616         path = strings.TrimSuffix(path, "/")
 617         var props map[string]interface{}
 618         for {
 619                 fi, err := fs.Stat(path)
 620                 if err != nil {
 621                         return err
 622                 }
 623                 switch src := fi.Sys().(type) {
 624                 case *arvados.Collection:
 625                         props = src.Properties
 626                 case *arvados.Group:
 627                         props = src.Properties
 628                 default:
 629                         if err, ok := src.(error); ok {
 630                                 return err
 631                         }
 632                         // Try parent
 633                         cut := strings.LastIndexByte(path, '/')
 634                         if cut < 0 {
 635                                 return nil
 636                         }
 637                         path = path[:cut]
 638                         continue
 639                 }
 640                 break
 641         }
 642         for k, v := range props {
 643                 if !validMIMEHeaderKey(k) {
 644                         continue
 645                 }
 646                 k = "x-amz-meta-" + k
 647                 if s, ok := v.(string); ok {
 648                         header.Set(k, maybeEncode(s))
 649                 } else if j, err := json.Marshal(v); err == nil {
 650                         header.Set(k, maybeEncode(string(j)))
 651                 }
 652         }
 653         return nil
 654 }
 655
 656 func validMIMEHeaderKey(k string) bool {
 657         check := "z-" + k
 658         return check != textproto.CanonicalMIMEHeaderKey(check)
 659 }
 660
 661 // Call fn on the given path (directory) and its contents, in
 662 // lexicographic order.
 663 //
 664 // If isRoot==true and path is not a directory, return nil.
 665 //
 666 // If fn returns filepath.SkipDir when called on a directory, don't
 667 // descend into that directory.
 668 func walkFS(fs arvados.CustomFileSystem, path string, isRoot bool, fn func(path string, fi os.FileInfo) error) error {
 669         if isRoot {
 670                 fi, err := fs.Stat(path)
 671                 if os.IsNotExist(err) || (err == nil && !fi.IsDir()) {
 672                         return nil
 673                 } else if err != nil {
 674                         return err
 675                 }
 676                 err = fn(path, fi)
 677                 if err == filepath.SkipDir {
 678                         return nil
 679                 } else if err != nil {
 680                         return err
 681                 }
 682         }
 683         f, err := fs.Open(path)
 684         if os.IsNotExist(err) && isRoot {
 685                 return nil
 686         } else if err != nil {
 687                 return fmt.Errorf("open %q: %w", path, err)
 688         }
 689         defer f.Close()
 690         if path == "/" {
 691                 path = ""
 692         }
 693         fis, err := f.Readdir(-1)
 694         if err != nil {
 695                 return err
 696         }
 697         sort.Slice(fis, func(i, j int) bool { return fis[i].Name() < fis[j].Name() })
 698         for _, fi := range fis {
 699                 err = fn(path+"/"+fi.Name(), fi)
 700                 if err == filepath.SkipDir {
 701                         continue
 702                 } else if err != nil {
 703                         return err
 704                 }
 705                 if fi.IsDir() {
 706                         err = walkFS(fs, path+"/"+fi.Name(), false, fn)
 707                         if err != nil {
 708                                 return err
 709                         }
 710                 }
 711         }
 712         return nil
 713 }
 714
 715 var errDone = errors.New("done")
 716
 717 func (h *handler) s3list(bucket string, w http.ResponseWriter, r *http.Request, fs arvados.CustomFileSystem) {
 718         var params struct {
 719                 v2                bool
 720                 delimiter         string
 721                 maxKeys           int
 722                 prefix            string
 723                 marker            string // decoded continuationToken (v2) or provided by client (v1)
 724                 startAfter        string // v2
 725                 continuationToken string // v2
 726                 encodingTypeURL   bool   // v2
 727         }
 728         params.delimiter = r.FormValue("delimiter")
 729         if mk, _ := strconv.ParseInt(r.FormValue("max-keys"), 10, 64); mk > 0 && mk < s3MaxKeys {
 730                 params.maxKeys = int(mk)
 731         } else {
 732                 params.maxKeys = s3MaxKeys
 733         }
 734         params.prefix = r.FormValue("prefix")
 735         switch r.FormValue("list-type") {
 736         case "":
 737         case "2":
 738                 params.v2 = true
 739         default:
 740                 http.Error(w, "invalid list-type parameter", http.StatusBadRequest)
 741                 return
 742         }
 743         if params.v2 {
 744                 params.continuationToken = r.FormValue("continuation-token")
 745                 marker, err := base64.StdEncoding.DecodeString(params.continuationToken)
 746                 if err != nil {
 747                         http.Error(w, "invalid continuation token", http.StatusBadRequest)
 748                         return
 749                 }
 750                 params.marker = string(marker)
 751                 params.startAfter = r.FormValue("start-after")
 752                 switch r.FormValue("encoding-type") {
 753                 case "":
 754                 case "url":
 755                         params.encodingTypeURL = true
 756                 default:
 757                         http.Error(w, "invalid encoding-type parameter", http.StatusBadRequest)
 758                         return
 759                 }
 760         } else {
 761                 params.marker = r.FormValue("marker")
 762         }
 763
 764         bucketdir := "by_id/" + bucket
 765         // walkpath is the directory (relative to bucketdir) we need
 766         // to walk: the innermost directory that is guaranteed to
 767         // contain all paths that have the requested prefix. Examples:
 768         // prefix "foo/bar"  => walkpath "foo"
 769         // prefix "foo/bar/" => walkpath "foo/bar"
 770         // prefix "foo"      => walkpath ""
 771         // prefix ""         => walkpath ""
 772         walkpath := params.prefix
 773         if cut := strings.LastIndex(walkpath, "/"); cut >= 0 {
 774                 walkpath = walkpath[:cut]
 775         } else {
 776                 walkpath = ""
 777         }
 778
 779         resp := listV2Resp{
 780                 Name:              bucket,
 781                 Prefix:            params.prefix,
 782                 Delimiter:         params.delimiter,
 783                 MaxKeys:           params.maxKeys,
 784                 ContinuationToken: r.FormValue("continuation-token"),
 785                 StartAfter:        params.startAfter,
 786         }
 787         nextMarker := ""
 788
 789         commonPrefixes := map[string]bool{}
 790         err := walkFS(fs, strings.TrimSuffix(bucketdir+"/"+walkpath, "/"), true, func(path string, fi os.FileInfo) error {
 791                 if path == bucketdir {
 792                         return nil
 793                 }
 794                 path = path[len(bucketdir)+1:]
 795                 filesize := fi.Size()
 796                 if fi.IsDir() {
 797                         path += "/"
 798                         filesize = 0
 799                 }
 800                 if len(path) <= len(params.prefix) {
 801                         if path > params.prefix[:len(path)] {
 802                                 // with prefix "foobar", walking "fooz" means we're done
 803                                 return errDone
 804                         }
 805                         if path < params.prefix[:len(path)] {
 806                                 // with prefix "foobar", walking "foobag" is pointless
 807                                 return filepath.SkipDir
 808                         }
 809                         if fi.IsDir() && !strings.HasPrefix(params.prefix+"/", path) {
 810                                 // with prefix "foo/bar", walking "fo"
 811                                 // is pointless (but walking "foo" or
 812                                 // "foo/bar" is necessary)
 813                                 return filepath.SkipDir
 814                         }
 815                         if len(path) < len(params.prefix) {
 816                                 // can't skip anything, and this entry
 817                                 // isn't in the results, so just
 818                                 // continue descent
 819                                 return nil
 820                         }
 821                 } else {
 822                         if path[:len(params.prefix)] > params.prefix {
 823                                 // with prefix "foobar", nothing we
 824                                 // see after "foozzz" is relevant
 825                                 return errDone
 826                         }
 827                 }
 828                 if path < params.marker || path < params.prefix || path <= params.startAfter {
 829                         return nil
 830                 }
 831                 if fi.IsDir() && !h.Cluster.Collections.S3FolderObjects {
 832                         // Note we don't add anything to
 833                         // commonPrefixes here even if delimiter is
 834                         // "/". We descend into the directory, and
 835                         // return a commonPrefix only if we end up
 836                         // finding a regular file inside it.
 837                         return nil
 838                 }
 839                 if len(resp.Contents)+len(commonPrefixes) >= params.maxKeys {
 840                         resp.IsTruncated = true
 841                         if params.delimiter != "" || params.v2 {
 842                                 nextMarker = path
 843                         }
 844                         return errDone
 845                 }
 846                 if params.delimiter != "" {
 847                         idx := strings.Index(path[len(params.prefix):], params.delimiter)
 848                         if idx >= 0 {
 849                                 // with prefix "foobar" and delimiter
 850                                 // "z", when we hit "foobar/baz", we
 851                                 // add "/baz" to commonPrefixes and
 852                                 // stop descending.
 853                                 commonPrefixes[path[:len(params.prefix)+idx+1]] = true
 854                                 return filepath.SkipDir
 855                         }
 856                 }
 857                 resp.Contents = append(resp.Contents, s3.Key{
 858                         Key:          path,
 859                         LastModified: fi.ModTime().UTC().Format("2006-01-02T15:04:05.999") + "Z",
 860                         Size:         filesize,
 861                 })
 862                 return nil
 863         })
 864         if err != nil && err != errDone {
 865                 http.Error(w, err.Error(), http.StatusInternalServerError)
 866                 return
 867         }
 868         if params.delimiter != "" {
 869                 resp.CommonPrefixes = make([]commonPrefix, 0, len(commonPrefixes))
 870                 for prefix := range commonPrefixes {
 871                         resp.CommonPrefixes = append(resp.CommonPrefixes, commonPrefix{prefix})
 872                 }
 873                 sort.Slice(resp.CommonPrefixes, func(i, j int) bool { return resp.CommonPrefixes[i].Prefix < resp.CommonPrefixes[j].Prefix })
 874         }
 875         resp.KeyCount = len(resp.Contents)
 876         var respV1orV2 interface{}
 877
 878         if params.encodingTypeURL {
 879                 // https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html
 880                 // "If you specify the encoding-type request
 881                 // parameter, Amazon S3 includes this element in the
 882                 // response, and returns encoded key name values in
 883                 // the following response elements:
 884                 //
 885                 // Delimiter, Prefix, Key, and StartAfter.
 886                 //
 887                 //      Type: String
 888                 //
 889                 // Valid Values: url"
 890                 //
 891                 // This is somewhat vague but in practice it appears
 892                 // to mean x-www-form-urlencoded as in RFC1866 8.2.1
 893                 // para 1 (encode space as "+") rather than straight
 894                 // percent-encoding as in RFC1738 2.2.  Presumably,
 895                 // the intent is to allow the client to decode XML and
 896                 // then paste the strings directly into another URI
 897                 // query or POST form like "https://host/path?foo=" +
 898                 // foo + "&bar=" + bar.
 899                 resp.EncodingType = "url"
 900                 resp.Delimiter = url.QueryEscape(resp.Delimiter)
 901                 resp.Prefix = url.QueryEscape(resp.Prefix)
 902                 resp.StartAfter = url.QueryEscape(resp.StartAfter)
 903                 for i, ent := range resp.Contents {
 904                         ent.Key = url.QueryEscape(ent.Key)
 905                         resp.Contents[i] = ent
 906                 }
 907                 for i, ent := range resp.CommonPrefixes {
 908                         ent.Prefix = url.QueryEscape(ent.Prefix)
 909                         resp.CommonPrefixes[i] = ent
 910                 }
 911         }
 912
 913         if params.v2 {
 914                 resp.NextContinuationToken = base64.StdEncoding.EncodeToString([]byte(nextMarker))
 915                 respV1orV2 = resp
 916         } else {
 917                 respV1orV2 = listV1Resp{
 918                         CommonPrefixes: resp.CommonPrefixes,
 919                         NextMarker:     nextMarker,
 920                         KeyCount:       resp.KeyCount,
 921                         ListResp: s3.ListResp{
 922                                 IsTruncated: resp.IsTruncated,
 923                                 Name:        bucket,
 924                                 Prefix:      params.prefix,
 925                                 Delimiter:   params.delimiter,
 926                                 Marker:      params.marker,
 927                                 MaxKeys:     params.maxKeys,
 928                                 Contents:    resp.Contents,
 929                         },
 930                 }
 931         }
 932
 933         w.Header().Set("Content-Type", "application/xml")
 934         io.WriteString(w, xml.Header)
 935         if err := xml.NewEncoder(w).Encode(respV1orV2); err != nil {
 936                 ctxlog.FromContext(r.Context()).WithError(err).Error("error writing xml response")
 937         }
 938 }